Interpretable multi-index framework for extracting dry-season built-up areas: benchmarking machine learning with Sentinel-2
Year: 2026
Type: article
Abstract: Spectral similarity between built-up surfaces and exposed dry soil significantly constrains built-up area extraction during dry seasons. Although machine-learning techniques can address this challenge by modeling complex spectral relationships, they generally require sufficiently large and representative labeled training datasets. This study therefore investigates a less data-demanding, rule-based multi-index approach for dry-season built-up mapping, aiming to reduce reliance on large labeled training datasets while maintaining effective classification performance. Sentinel-2A imagery acquired in April 2024 was analyzed, and temporal validation was conducted for 2020, 2022, and 2024. Among the tested combinations, the Built-up Area Extraction Index (BAEI), Dry BareSoil Index (DBSI), and Normalized Difference Vegetation Index (NDVI) achieved the highest performance, yielding 95 % overall accuracy and a Kappa coefficient of 0.89. This represents a substantial improvement over singleindex BAEI (80 % overall accuracy; Kappa 0.55), increasing the built-up user’s accuracy from 61 % to 92 %. Multitemporal validation confirmed that the optimized BAEI–DBSI–NDVI decision rules remained stable across the 2020–2024 dry-season images without recalibration, achieving 93–96 % overall accuracy and a built-up F1-score of 84–93 %. Spatial analysis demonstrated robust performance across peripheral and agricultural zones, with moderate variability in dense urban cores due to spectral heterogeneity. To evaluate the operational robustness of the proposed framework, it was benchmarked against machine-learning classifiers, specifically Support Vector Machine and Random Forest (RF). Under the specific conditions of this study, the proposed rule-based method (95 % accuracy) marginally outperformed both SVM (93 %) and RF (92 %), while offering greater transparency and reducing dependency on large, labeled training datasets. Furthermore, feature importance analysis confirmed the critical role of these selected indices in resolving spectral confusion. These findings suggest that the proposed framework offers a transparent and computationally efficient approach for dry-season urban monitoring in tropical coastal environments similar to Visakhapatnam.
Authors: Sarah Sejari, Vazeer Mahammood
Institution: Andhra University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7214385607